| name | med-autoscience |
| description | Use when Codex needs MedAutoScience (MAS) to plan, conduct, review, or publish medical research, including research questions, study design, cohorts and endpoints, evidence synthesis, bounded analysis, scientific figures/tables, manuscript writing, quality review, and publication handoff. Do not use for patient-specific diagnosis, treatment, triage, or emergency advice; medical grant applications; or generic Office formatting without a research objective. |
MedAutoScience
Canonical OPL Agent id is mas; med-autoscience is the package/plugin Skill locator. Operate MAS through OPL-generated interfaces while MAS retains medical, evidence, artifact, quality, and publication authority.
Admission
- Admit MAS only when the requested outcome is medical research or a research artifact tied to an identifiable study, evidence question, dataset/cohort, manuscript, or publication package.
- Do not route patient-specific clinical-care requests to MAS. Keep diagnosis, treatment choice, dosing, triage, and emergency guidance outside this research Agent and direct the user to appropriate clinical care when needed.
- Route funding-call strategy and grant application authoring to MAG. Route generic slides to RCA and generic document formatting to the relevant document capability unless the work is subordinate to a MAS research artifact.
- Bind the current research objective, source/data cohort, study/workspace identity, accepted upstream refs, and intended claim before selecting an action. Do not infer scientific readiness from file presence or package/runtime status.
Action Routing
Choose the earliest unresolved owning Stage action from the installed OPL-generated interface:
direction_and_route_selection: define or repair the research question, route, feasibility, and governing assumptions.
baseline_and_evidence_setup: establish the protocol/method baseline, evidence/source map, cohort/phenotype, endpoints, governance, and analysis readiness.
bounded_analysis_campaign: execute a bounded, reproducible analysis against an accepted design and frozen inputs.
manuscript_authoring: write or materially revise the scientific manuscript and its display artifacts from current evidence refs.
review_and_quality_gate: run independent scientific, statistical, evidence-integrity, reference, figure/table, or manuscript review.
finalize_and_publication_handoff: assemble the final publication handoff after current review evidence and explicit human submission authority.
qualification_work_item_provisioning_authority_evaluate, study_lifecycle_reactivation_authority_evaluate, candidate_admission_authority_evaluate, build_dependency_currentness_authority_evaluate, and paper_mission_authority_evaluate are internal registry-bound authority actions. They have no public CLI, MCP, Skill, product-entry, OpenAI, or AI SDK route and must not be invoked as user actions.
Default Workflow
- Select one public action from the research intent and current accepted refs; do not begin with package lifecycle or environment commands when the user asked for research work.
- For an inactive study receiving an explicit manuscript-revision instruction, first persist durable user-instruction evidence and a
reviewer_revision intake that binds the full checklist, independent review packet, first owning Stage, and allowed revision scope. Use the hosted lifecycle admission flow; runtime activity alone never reactivates MAS business truth.
- Preserve one OPL StageRun lineage and exact source/data/artifact refs while Codex routes among declared MAS stages.
- Use the installed
mas-scholar-skills capability package for focused medical methods. Acquire another external Skill only for a named or demonstrated coverage gap after provenance, permissions, data/credential scope, and compatibility review.
- For every initial draft, apply the MAS-owned
initial_draft_evidence_integrity_requirements in contracts/manuscript_first_draft_quality_policy.json at their declared earliest Stage owners. Missing or unresolved requirements are quality debt and cannot be hidden by prose, Skill output, render success, or runtime completion.
- Write as the manuscript authors. Route scientific evidence gaps back, but preserve the ideal manuscript structure for author-supplied objective facts and insert the minimum local
[AUTHOR INPUT: ...] annotation. Bind every annotation and its derived To-Do projection through one author_stance_integrity_ref; pending author facts never justify defensive workflow prose.
- Prepare the declarative runtime environment only when the selected analysis/display work requires it. An environment receipt proves execution readiness only, never medical or publication quality.
- Separate producing and independent reviewer/auditor invocations, task records, and receipts before closing a formal quality gate.
- Freeze one candidate after bounded previews, then run complete export, exact-byte inventory, and affected independent review once for that candidate.
Quality And Hard Stops
- Treat raw/partial artifacts, negative or null results, failed attempts, review findings, and no-output diagnostics as route context. A consumable delta may advance as
completed_with_quality_debt.
- Do not edit
submission/ or launch a Stage from a paused, delivered-paused, or stopped business projection. Reactivation requires a current MAS receipt plus matching OPL all-or-rollback materialization receipt; neither a Temporal-running signal nor provider completion is sufficient.
- A qualification-only work item may prove Full VM identity and lifecycle transport only. It never authorizes a Stage body, business action, publication, or submission, even after its MAS provisioning receipt and matching OPL materialization receipt are current.
- Retry, review, and repair limits are quality budgets. Exhaustion preserves the best evidence-bearing artifact and closes quality, publication, export, and submission-ready claims; it does not create an infinite stage loop.
- Negative findings remain evidence. Preserve lineage and route back to the owning hypothesis, cohort/data, endpoint, analysis design, writing, or review stage instead of optimizing for a positive result.
- For new or materially repaired paper-facing figures, use
medical-figure-design; use medical-figure-style for final visual QA and medical-figure-composer only for separately assembled panels.
- Only executor unavailability, wrong-target identity/currentness, real authority/safety/credential boundaries, irreversible action, protected data/material, or an explicit human decision may hard-stop progress.
Output Expectations
- Return the selected public action and Stage, exact research/source/data/artifact refs, methods or analysis outputs, review/authority refs, typed blockers, remaining quality debt, and the next owning Stage.
- Match every scientific claim to its evidence class and locator. Do not promote external Skill output, preview/cache evidence, runtime receipts, generated descriptors, or provider completion into MAS authority.
- Name artifact maturity precisely: plan, frozen baseline, bounded analysis result, manuscript draft, review finding, publication handoff candidate, or owner-authorized final handoff.
- Preserve MAS-owned owner receipts and publication/submission authority; OPL may transport and persist refs but cannot author medical truth or sign those decisions.
References
contracts/action_catalog.json
agent/stages/manifest.json
contracts/artifact_iteration_efficiency_policy.json
contracts/stage_quality_cycle_policy.json
contracts/runtime_environment_requirements.json
contracts/owner_receipt_contract.json
contracts/domain_handler_registry.json